On Exploring Shape and Semantic Enhancements for RGB-X Semantic Segmentation